Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * X86-64 specific CPU setup. |
| 3 | * Copyright (C) 1995 Linus Torvalds |
| 4 | * Copyright 2001, 2002, 2003 SuSE Labs / Andi Kleen. |
| 5 | * See setup.c for older changelog. |
| 6 | * $Id: setup64.c,v 1.12 2002/03/21 10:09:17 ak Exp $ |
| 7 | */ |
| 8 | #include <linux/config.h> |
| 9 | #include <linux/init.h> |
| 10 | #include <linux/kernel.h> |
| 11 | #include <linux/sched.h> |
| 12 | #include <linux/string.h> |
| 13 | #include <linux/bootmem.h> |
| 14 | #include <linux/bitops.h> |
| 15 | #include <asm/pda.h> |
| 16 | #include <asm/pgtable.h> |
| 17 | #include <asm/processor.h> |
| 18 | #include <asm/desc.h> |
| 19 | #include <asm/atomic.h> |
| 20 | #include <asm/mmu_context.h> |
| 21 | #include <asm/smp.h> |
| 22 | #include <asm/i387.h> |
| 23 | #include <asm/percpu.h> |
| 24 | #include <asm/mtrr.h> |
| 25 | #include <asm/proto.h> |
| 26 | #include <asm/mman.h> |
| 27 | #include <asm/numa.h> |
| 28 | |
| 29 | char x86_boot_params[2048] __initdata = {0,}; |
| 30 | |
| 31 | cpumask_t cpu_initialized __initdata = CPU_MASK_NONE; |
| 32 | |
| 33 | struct x8664_pda cpu_pda[NR_CPUS] __cacheline_aligned; |
| 34 | |
| 35 | extern struct task_struct init_task; |
| 36 | |
| 37 | extern unsigned char __per_cpu_start[], __per_cpu_end[]; |
| 38 | |
| 39 | extern struct desc_ptr cpu_gdt_descr[]; |
| 40 | struct desc_ptr idt_descr = { 256 * 16, (unsigned long) idt_table }; |
| 41 | |
| 42 | char boot_cpu_stack[IRQSTACKSIZE] __attribute__((section(".bss.page_aligned"))); |
| 43 | |
| 44 | unsigned long __supported_pte_mask = ~0UL; |
| 45 | static int do_not_nx __initdata = 0; |
| 46 | |
| 47 | /* noexec=on|off |
| 48 | Control non executable mappings for 64bit processes. |
| 49 | |
| 50 | on Enable(default) |
| 51 | off Disable |
| 52 | */ |
| 53 | int __init nonx_setup(char *str) |
| 54 | { |
| 55 | if (!strncmp(str, "on", 2)) { |
| 56 | __supported_pte_mask |= _PAGE_NX; |
| 57 | do_not_nx = 0; |
| 58 | } else if (!strncmp(str, "off", 3)) { |
| 59 | do_not_nx = 1; |
| 60 | __supported_pte_mask &= ~_PAGE_NX; |
| 61 | } |
| 62 | return 0; |
| 63 | } |
| 64 | __setup("noexec=", nonx_setup); /* parsed early actually */ |
| 65 | |
| 66 | int force_personality32 = READ_IMPLIES_EXEC; |
| 67 | |
| 68 | /* noexec32=on|off |
| 69 | Control non executable heap for 32bit processes. |
| 70 | To control the stack too use noexec=off |
| 71 | |
| 72 | on PROT_READ does not imply PROT_EXEC for 32bit processes |
| 73 | off PROT_READ implies PROT_EXEC (default) |
| 74 | */ |
| 75 | static int __init nonx32_setup(char *str) |
| 76 | { |
| 77 | if (!strcmp(str, "on")) |
| 78 | force_personality32 &= ~READ_IMPLIES_EXEC; |
| 79 | else if (!strcmp(str, "off")) |
| 80 | force_personality32 |= READ_IMPLIES_EXEC; |
| 81 | return 0; |
| 82 | } |
| 83 | __setup("noexec32=", nonx32_setup); |
| 84 | |
| 85 | /* |
| 86 | * Great future plan: |
| 87 | * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data. |
| 88 | * Always point %gs to its beginning |
| 89 | */ |
| 90 | void __init setup_per_cpu_areas(void) |
| 91 | { |
| 92 | int i; |
| 93 | unsigned long size; |
| 94 | |
| 95 | /* Copy section for each CPU (we discard the original) */ |
| 96 | size = ALIGN(__per_cpu_end - __per_cpu_start, SMP_CACHE_BYTES); |
| 97 | #ifdef CONFIG_MODULES |
| 98 | if (size < PERCPU_ENOUGH_ROOM) |
| 99 | size = PERCPU_ENOUGH_ROOM; |
| 100 | #endif |
| 101 | |
| 102 | for (i = 0; i < NR_CPUS; i++) { |
| 103 | unsigned char *ptr; |
| 104 | |
| 105 | if (!NODE_DATA(cpu_to_node(i))) { |
| 106 | printk("cpu with no node %d, num_online_nodes %d\n", |
| 107 | i, num_online_nodes()); |
| 108 | ptr = alloc_bootmem(size); |
| 109 | } else { |
| 110 | ptr = alloc_bootmem_node(NODE_DATA(cpu_to_node(i)), size); |
| 111 | } |
| 112 | if (!ptr) |
| 113 | panic("Cannot allocate cpu data for CPU %d\n", i); |
| 114 | cpu_pda[i].data_offset = ptr - __per_cpu_start; |
| 115 | memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start); |
| 116 | } |
| 117 | } |
| 118 | |
| 119 | void pda_init(int cpu) |
| 120 | { |
| 121 | struct x8664_pda *pda = &cpu_pda[cpu]; |
| 122 | |
| 123 | /* Setup up data that may be needed in __get_free_pages early */ |
| 124 | asm volatile("movl %0,%%fs ; movl %0,%%gs" :: "r" (0)); |
| 125 | wrmsrl(MSR_GS_BASE, cpu_pda + cpu); |
| 126 | |
| 127 | pda->me = pda; |
| 128 | pda->cpunumber = cpu; |
| 129 | pda->irqcount = -1; |
| 130 | pda->kernelstack = |
| 131 | (unsigned long)stack_thread_info() - PDA_STACKOFFSET + THREAD_SIZE; |
| 132 | pda->active_mm = &init_mm; |
| 133 | pda->mmu_state = 0; |
| 134 | |
| 135 | if (cpu == 0) { |
| 136 | /* others are initialized in smpboot.c */ |
| 137 | pda->pcurrent = &init_task; |
| 138 | pda->irqstackptr = boot_cpu_stack; |
| 139 | } else { |
| 140 | pda->irqstackptr = (char *) |
| 141 | __get_free_pages(GFP_ATOMIC, IRQSTACK_ORDER); |
| 142 | if (!pda->irqstackptr) |
| 143 | panic("cannot allocate irqstack for cpu %d", cpu); |
| 144 | } |
| 145 | |
| 146 | asm volatile("movq %0,%%cr3" :: "r" (__pa_symbol(&init_level4_pgt))); |
| 147 | |
| 148 | pda->irqstackptr += IRQSTACKSIZE-64; |
| 149 | } |
| 150 | |
| 151 | char boot_exception_stacks[N_EXCEPTION_STACKS * EXCEPTION_STKSZ] |
| 152 | __attribute__((section(".bss.page_aligned"))); |
| 153 | |
| 154 | /* May not be marked __init: used by software suspend */ |
| 155 | void syscall_init(void) |
| 156 | { |
| 157 | /* |
| 158 | * LSTAR and STAR live in a bit strange symbiosis. |
| 159 | * They both write to the same internal register. STAR allows to set CS/DS |
| 160 | * but only a 32bit target. LSTAR sets the 64bit rip. |
| 161 | */ |
| 162 | wrmsrl(MSR_STAR, ((u64)__USER32_CS)<<48 | ((u64)__KERNEL_CS)<<32); |
| 163 | wrmsrl(MSR_LSTAR, system_call); |
| 164 | |
| 165 | #ifdef CONFIG_IA32_EMULATION |
| 166 | syscall32_cpu_init (); |
| 167 | #endif |
| 168 | |
| 169 | /* Flags to clear on syscall */ |
| 170 | wrmsrl(MSR_SYSCALL_MASK, EF_TF|EF_DF|EF_IE|0x3000); |
| 171 | } |
| 172 | |
| 173 | void __init check_efer(void) |
| 174 | { |
| 175 | unsigned long efer; |
| 176 | |
| 177 | rdmsrl(MSR_EFER, efer); |
| 178 | if (!(efer & EFER_NX) || do_not_nx) { |
| 179 | __supported_pte_mask &= ~_PAGE_NX; |
| 180 | } |
| 181 | } |
| 182 | |
| 183 | /* |
| 184 | * cpu_init() initializes state that is per-CPU. Some data is already |
| 185 | * initialized (naturally) in the bootstrap process, such as the GDT |
| 186 | * and IDT. We reload them nevertheless, this function acts as a |
| 187 | * 'CPU state barrier', nothing should get across. |
| 188 | * A lot of state is already set up in PDA init. |
| 189 | */ |
| 190 | void __init cpu_init (void) |
| 191 | { |
| 192 | #ifdef CONFIG_SMP |
| 193 | int cpu = stack_smp_processor_id(); |
| 194 | #else |
| 195 | int cpu = smp_processor_id(); |
| 196 | #endif |
| 197 | struct tss_struct *t = &per_cpu(init_tss, cpu); |
| 198 | unsigned long v; |
| 199 | char *estacks = NULL; |
| 200 | struct task_struct *me; |
| 201 | int i; |
| 202 | |
| 203 | /* CPU 0 is initialised in head64.c */ |
| 204 | if (cpu != 0) { |
| 205 | pda_init(cpu); |
| 206 | } else |
| 207 | estacks = boot_exception_stacks; |
| 208 | |
| 209 | me = current; |
| 210 | |
| 211 | if (cpu_test_and_set(cpu, cpu_initialized)) |
| 212 | panic("CPU#%d already initialized!\n", cpu); |
| 213 | |
| 214 | printk("Initializing CPU#%d\n", cpu); |
| 215 | |
| 216 | clear_in_cr4(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE); |
| 217 | |
| 218 | /* |
| 219 | * Initialize the per-CPU GDT with the boot GDT, |
| 220 | * and set up the GDT descriptor: |
| 221 | */ |
| 222 | if (cpu) { |
| 223 | memcpy(cpu_gdt_table[cpu], cpu_gdt_table[0], GDT_SIZE); |
| 224 | } |
| 225 | |
| 226 | cpu_gdt_descr[cpu].size = GDT_SIZE; |
| 227 | cpu_gdt_descr[cpu].address = (unsigned long)cpu_gdt_table[cpu]; |
| 228 | asm volatile("lgdt %0" :: "m" (cpu_gdt_descr[cpu])); |
| 229 | asm volatile("lidt %0" :: "m" (idt_descr)); |
| 230 | |
| 231 | memcpy(me->thread.tls_array, cpu_gdt_table[cpu], GDT_ENTRY_TLS_ENTRIES * 8); |
| 232 | |
| 233 | /* |
| 234 | * Delete NT |
| 235 | */ |
| 236 | |
| 237 | asm volatile("pushfq ; popq %%rax ; btr $14,%%rax ; pushq %%rax ; popfq" ::: "eax"); |
| 238 | |
| 239 | syscall_init(); |
| 240 | |
| 241 | wrmsrl(MSR_FS_BASE, 0); |
| 242 | wrmsrl(MSR_KERNEL_GS_BASE, 0); |
| 243 | barrier(); |
| 244 | |
| 245 | check_efer(); |
| 246 | |
| 247 | /* |
| 248 | * set up and load the per-CPU TSS |
| 249 | */ |
| 250 | for (v = 0; v < N_EXCEPTION_STACKS; v++) { |
| 251 | if (cpu) { |
| 252 | estacks = (char *)__get_free_pages(GFP_ATOMIC, |
| 253 | EXCEPTION_STACK_ORDER); |
| 254 | if (!estacks) |
| 255 | panic("Cannot allocate exception stack %ld %d\n", |
| 256 | v, cpu); |
| 257 | } |
| 258 | estacks += EXCEPTION_STKSZ; |
| 259 | t->ist[v] = (unsigned long)estacks; |
| 260 | } |
| 261 | |
| 262 | t->io_bitmap_base = offsetof(struct tss_struct, io_bitmap); |
| 263 | /* |
| 264 | * <= is required because the CPU will access up to |
| 265 | * 8 bits beyond the end of the IO permission bitmap. |
| 266 | */ |
| 267 | for (i = 0; i <= IO_BITMAP_LONGS; i++) |
| 268 | t->io_bitmap[i] = ~0UL; |
| 269 | |
| 270 | atomic_inc(&init_mm.mm_count); |
| 271 | me->active_mm = &init_mm; |
| 272 | if (me->mm) |
| 273 | BUG(); |
| 274 | enter_lazy_tlb(&init_mm, me); |
| 275 | |
| 276 | set_tss_desc(cpu, t); |
| 277 | load_TR_desc(); |
| 278 | load_LDT(&init_mm.context); |
| 279 | |
| 280 | /* |
| 281 | * Clear all 6 debug registers: |
| 282 | */ |
| 283 | |
| 284 | set_debug(0UL, 0); |
| 285 | set_debug(0UL, 1); |
| 286 | set_debug(0UL, 2); |
| 287 | set_debug(0UL, 3); |
| 288 | set_debug(0UL, 6); |
| 289 | set_debug(0UL, 7); |
| 290 | |
| 291 | fpu_init(); |
| 292 | } |